Qwen3-8B-9c925d64-bf16-GGUF

Qwen3-8B — bf16 GGUF for TensorCash mining

Qwen3-8B-9c925d64-bf16.gguf — the model used by the TensorCash (TSC) mainnet proof-of-inference chain, converted to GGUF in bf16 (no quantisation).

Why bf16 and not a quant

A TSC share is a 256-token sampling transcript that a verifier replays. The consensus rule snaps logits to bf16, so the transcript only reproduces if the weights are bit-identical to the registered model. Any quantisation (Q8_0, Q4_K, …) changes the logits and every share is rejected. Use this file as-is.

• size: 16,388,043,744 bytes
• precision: bf16
• vocabulary: 151,936

Provenance

Converted with llama.cpp/convert_hf_to_gguf.py --outtype bf16 from the Hugging Face Qwen3-8B revision registered by the chain (commit prefix 9c925d64, which is where the filename suffix comes from). Tensor-by-tensor comparison against an independently produced conversion matched on 64/64 tensors.

Use

docker run --gpus all \
  -e POOL_URL=stratum+tcp://POOL:PORT \
  -e WALLET=<your address> -e WORKER=rig1 \
  -e MODEL_URL=https://huggingface.co/ocminer/Qwen3-8B-9c925d64-bf16-GGUF/resolve/main/Qwen3-8B-9c925d64-bf16.gguf \
  ocminersupr/supr-meow-tsc:latest

The miner (ocminer/supr-meow-tsc) auto-detects the GPU and loads a measured slots/groups profile.

Licensed Apache-2.0, following the upstream Qwen3-8B release.